🦴 Osteoporosis X‑Ray Analyzer — Science Fair Edition

A student-friendly project that demonstrates how computer vision (AI) and simple mechanical measurements can work together to estimate bone weakening on humerus X‑rays. It runs as a small web app on your laptop—no command line needed.

What this project does

• Classifies an input X‑ray as “Normal” or “Osteoporotic” using a pre‑trained EfficientNet‑B0 model
• Combines the AI score with lab measurements (fracture force and propagation speed) to compute a Severity score (0–1)
• Converts the score into an easy category: Normal, Mild, Moderate, Severe
• Shows a friendly recommendation message and lets you download a text report

Educational use only. This is NOT a medical device and must not be used for diagnosis or treatment decisions.

---

Materials required (for demo at the fair)

• A Windows laptop (8 GB RAM or more; GPU optional)
• This project folder
• Python 3.10+ installed
• A couple hundred sample humerus X‑ray images (JPG/PNG)
• The bundled CSV with lab test values: humpty dumpty is humping my leg.csv

Optional (for advanced experimentation):

• NVIDIA GPU with CUDA for faster inference (not required)
• Access to the MURA dataset if you want to reproduce training

---

Software dependencies

The app uses these Python packages (see requirements.txt):

• torch
• torchvision
• Pillow
• timm
• streamlit

Install them once:

pip install -r requirements.txt

---

Training data and model

• Base dataset: MURA (Musculoskeletal Radiographs) — publicly available research dataset by Stanford ML Group.
• Subset used: Humerus X‑rays. Studies labeled “positive” (abnormal) vs “negative” (normal). For this demo, “positive” is treated as osteoporotic, which is a simplification for learning purposes.
• Model: EfficientNet‑B0 fine‑tuned for 2 classes (Normal vs Osteoporotic). The trained weights are already included as efficientnet_humerus.pt.

Limitation: MURA’s “abnormal” label is not a clinical osteoporosis label. This project demonstrates technique, not clinical diagnostic accuracy.

---

How it works (under the hood)

1. Image analysis (AI):

   • Preprocess the X‑ray to 224×224, normalize, and run it through EfficientNet‑B0.
   • Output probabilities: Normal and Osteoporotic.

2. Lab test features (from CSV):

   • Breaking Point (lbs) → converted to Newtons: 1 lbf = 4.4482216153 N.
   • Fracture Propagation Speed (m/s) → converted to mm/s.
   • The app computes reference values from the CSV:
     • F_REF = mean(failure_load_n)
     • V_REF = mean(prop_speed_mm_s)
     • V_MAX = max(prop_speed_mm_s) (with a small safety headroom)

3. Severity score (no re‑training needed):

   S_F = clamp(failure_load_n / F_REF, 0, 1)
   S_v = clamp((prop_speed_mm_s − V_REF) / (V_MAX − V_REF), 0, 1)
   severity = 0.45 * model_prob_osteo + 0.35 * (1 − S_F) + 0.20 * S_v
   

4. Category + recommendation:

   • Normal (< 0.30) — Mild (0.30–0.55) — Moderate (0.55–0.75) — Severe (≥ 0.75)
   • Human‑readable message appears with a soft severity‑colored card.

---

How to run the app (no command line required)

Option 1 — Double‑click

• Double‑click run_app.bat
• Your browser opens to http://localhost:8501

Option 2 — From PowerShell

pip install -r requirements.txt   # only once
streamlit run app.py

In the app:

• In the sidebar: select a Trial from the bundled CSV. The app calculates reference values and fills in the trial’s measurements (with friendly unit conversions).
• In the main panel: upload an X‑ray image and click “Analyze X‑Ray”.
• You’ll see prediction, probabilities, a severity category card, and a message. You can download a text report.

Stopping the app:

• Press Ctrl + C in the Streamlit terminal window to stop the server.

---

What it can and cannot do

What it can do

• Demonstrate an end‑to‑end AI pipeline for medical‑style images
• Show how to combine model probabilities with mechanical measurements
• Convert real‑world units (lb→N, m/s→mm/s) and normalize against references
• Provide a simple severity score and friendly recommendation message

What it cannot do

• Provide a medical diagnosis (it’s for learning only)
• Guarantee accuracy on all bones/images/settings
• Replace qualified clinical assessment or lab testing
• Infer personalized risk without proper clinical context

---

Reproduce training (optional)

If you want to explore the dataset and training flow:

1. Prepare sorted data from MURA Humerus studies (positive/negative):
   • Edit RAW_ROOT in sort_mura_humerus.py to your MURA path
   • Run it to populate dataset_sorted/normal and dataset_sorted/osteoporotic
2. Train EfficientNet‑B0 on your machine:
   • Run python train.py (tweaks: batch size, epochs, LR)
   • The model is saved as efficientnet_humerus.pt

Training is optional; the app already includes a trained model.

---

Ethics and safety

• Do not use this tool for medical decisions.

---

Credits

• MURA Dataset — Stanford ML Group
• EfficientNet — Tan & Le
• Streamlit — the web UI framework used here